1 // SPDX-License-Identifier: GPL-2.0
3 * bcachefs journalling code, for btree insertions
5 * Copyright 2012 Google, Inc.
9 #include "alloc_foreground.h"
10 #include "bkey_methods.h"
14 #include "journal_io.h"
15 #include "journal_reclaim.h"
16 #include "journal_seq_blacklist.h"
19 #include <trace/events/bcachefs.h>
21 static bool __journal_entry_is_open(union journal_res_state state)
23 return state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL;
26 static bool journal_entry_is_open(struct journal *j)
28 return __journal_entry_is_open(j->reservations);
31 static void journal_pin_new_entry(struct journal *j, int count)
33 struct journal_entry_pin_list *p;
36 * The fifo_push() needs to happen at the same time as j->seq is
37 * incremented for journal_last_seq() to be calculated correctly
39 atomic64_inc(&j->seq);
40 p = fifo_push_ref(&j->pin);
42 INIT_LIST_HEAD(&p->list);
43 INIT_LIST_HEAD(&p->flushed);
44 atomic_set(&p->count, count);
48 static void bch2_journal_buf_init(struct journal *j)
50 struct journal_buf *buf = journal_cur_buf(j);
52 memset(buf->has_inode, 0, sizeof(buf->has_inode));
54 memset(buf->data, 0, sizeof(*buf->data));
55 buf->data->seq = cpu_to_le64(journal_cur_seq(j));
59 void bch2_journal_halt(struct journal *j)
61 union journal_res_state old, new;
62 u64 v = atomic64_read(&j->reservations.counter);
66 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
69 new.cur_entry_offset = JOURNAL_ENTRY_ERROR_VAL;
70 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
71 old.v, new.v)) != old.v);
74 closure_wake_up(&journal_cur_buf(j)->wait);
77 /* journal entry close/open: */
79 void __bch2_journal_buf_put(struct journal *j, bool need_write_just_set)
81 if (!need_write_just_set &&
82 test_bit(JOURNAL_NEED_WRITE, &j->flags))
83 bch2_time_stats_update(j->delay_time,
86 clear_bit(JOURNAL_NEED_WRITE, &j->flags);
88 closure_call(&j->io, bch2_journal_write, system_highpri_wq, NULL);
92 * Returns true if journal entry is now closed:
94 static bool __journal_entry_close(struct journal *j)
96 struct bch_fs *c = container_of(j, struct bch_fs, journal);
97 struct journal_buf *buf = journal_cur_buf(j);
98 union journal_res_state old, new;
99 u64 v = atomic64_read(&j->reservations.counter);
100 bool set_need_write = false;
103 lockdep_assert_held(&j->lock);
107 if (old.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL)
110 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL) {
111 /* this entry will never be written: */
112 closure_wake_up(&buf->wait);
116 if (!test_bit(JOURNAL_NEED_WRITE, &j->flags)) {
117 set_bit(JOURNAL_NEED_WRITE, &j->flags);
118 j->need_write_time = local_clock();
119 set_need_write = true;
122 if (new.prev_buf_unwritten)
125 new.cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL;
127 new.prev_buf_unwritten = 1;
129 BUG_ON(journal_state_count(new, new.idx));
130 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
131 old.v, new.v)) != old.v);
133 buf->data->u64s = cpu_to_le32(old.cur_entry_offset);
135 sectors = vstruct_blocks_plus(buf->data, c->block_bits,
136 buf->u64s_reserved) << c->block_bits;
137 BUG_ON(sectors > buf->sectors);
138 buf->sectors = sectors;
140 bkey_extent_init(&buf->key);
143 * We have to set last_seq here, _before_ opening a new journal entry:
145 * A threads may replace an old pin with a new pin on their current
146 * journal reservation - the expectation being that the journal will
147 * contain either what the old pin protected or what the new pin
150 * After the old pin is dropped journal_last_seq() won't include the old
151 * pin, so we can only write the updated last_seq on the entry that
152 * contains whatever the new pin protects.
154 * Restated, we can _not_ update last_seq for a given entry if there
155 * could be a newer entry open with reservations/pins that have been
158 * Hence, we want update/set last_seq on the current journal entry right
159 * before we open a new one:
161 buf->data->last_seq = cpu_to_le64(journal_last_seq(j));
163 if (journal_entry_empty(buf->data))
164 clear_bit(JOURNAL_NOT_EMPTY, &j->flags);
166 set_bit(JOURNAL_NOT_EMPTY, &j->flags);
168 journal_pin_new_entry(j, 1);
170 bch2_journal_buf_init(j);
172 cancel_delayed_work(&j->write_work);
174 bch2_journal_space_available(j);
176 bch2_journal_buf_put(j, old.idx, set_need_write);
180 static bool journal_entry_close(struct journal *j)
185 ret = __journal_entry_close(j);
186 spin_unlock(&j->lock);
192 * should _only_ called from journal_res_get() - when we actually want a
193 * journal reservation - journal entry is open means journal is dirty:
197 * -ENOSPC: journal currently full, must invoke reclaim
198 * -EAGAIN: journal blocked, must wait
199 * -EROFS: insufficient rw devices or journal error
201 static int journal_entry_open(struct journal *j)
203 struct journal_buf *buf = journal_cur_buf(j);
204 union journal_res_state old, new;
208 lockdep_assert_held(&j->lock);
209 BUG_ON(journal_entry_is_open(j));
214 if (j->cur_entry_error)
215 return j->cur_entry_error;
217 BUG_ON(!j->cur_entry_sectors);
219 buf->u64s_reserved = j->entry_u64s_reserved;
220 buf->disk_sectors = j->cur_entry_sectors;
221 buf->sectors = min(buf->disk_sectors, buf->buf_size >> 9);
223 u64s = (int) (buf->sectors << 9) / sizeof(u64) -
224 journal_entry_overhead(j);
225 u64s = clamp_t(int, u64s, 0, JOURNAL_ENTRY_CLOSED_VAL - 1);
227 if (u64s <= le32_to_cpu(buf->data->u64s))
231 * Must be set before marking the journal entry as open:
233 j->cur_entry_u64s = u64s;
235 v = atomic64_read(&j->reservations.counter);
239 if (old.cur_entry_offset == JOURNAL_ENTRY_ERROR_VAL)
242 /* Handle any already added entries */
243 new.cur_entry_offset = le32_to_cpu(buf->data->u64s);
245 EBUG_ON(journal_state_count(new, new.idx));
246 journal_state_inc(&new);
247 } while ((v = atomic64_cmpxchg(&j->reservations.counter,
248 old.v, new.v)) != old.v);
250 if (j->res_get_blocked_start)
251 bch2_time_stats_update(j->blocked_time,
252 j->res_get_blocked_start);
253 j->res_get_blocked_start = 0;
255 mod_delayed_work(system_freezable_wq,
257 msecs_to_jiffies(j->write_delay_ms));
262 static bool journal_quiesced(struct journal *j)
264 union journal_res_state state = READ_ONCE(j->reservations);
265 bool ret = !state.prev_buf_unwritten && !__journal_entry_is_open(state);
268 journal_entry_close(j);
272 static void journal_quiesce(struct journal *j)
274 wait_event(j->wait, journal_quiesced(j));
277 static void journal_write_work(struct work_struct *work)
279 struct journal *j = container_of(work, struct journal, write_work.work);
281 journal_entry_close(j);
285 * Given an inode number, if that inode number has data in the journal that
286 * hasn't yet been flushed, return the journal sequence number that needs to be
289 u64 bch2_inode_journal_seq(struct journal *j, u64 inode)
291 size_t h = hash_64(inode, ilog2(sizeof(j->buf[0].has_inode) * 8));
294 if (!test_bit(h, j->buf[0].has_inode) &&
295 !test_bit(h, j->buf[1].has_inode))
299 if (test_bit(h, journal_cur_buf(j)->has_inode))
300 seq = journal_cur_seq(j);
301 else if (test_bit(h, journal_prev_buf(j)->has_inode))
302 seq = journal_cur_seq(j) - 1;
303 spin_unlock(&j->lock);
308 static int __journal_res_get(struct journal *j, struct journal_res *res,
311 struct bch_fs *c = container_of(j, struct bch_fs, journal);
312 struct journal_buf *buf;
316 if (journal_res_get_fast(j, res, flags))
319 if (bch2_journal_error(j))
325 * Recheck after taking the lock, so we don't race with another thread
326 * that just did journal_entry_open() and call journal_entry_close()
329 if (journal_res_get_fast(j, res, flags)) {
330 spin_unlock(&j->lock);
334 if (!(flags & JOURNAL_RES_GET_RESERVED) &&
335 !test_bit(JOURNAL_MAY_GET_UNRESERVED, &j->flags)) {
337 * Don't want to close current journal entry, just need to
345 * If we couldn't get a reservation because the current buf filled up,
346 * and we had room for a bigger entry on disk, signal that we want to
347 * realloc the journal bufs:
349 buf = journal_cur_buf(j);
350 if (journal_entry_is_open(j) &&
351 buf->buf_size >> 9 < buf->disk_sectors &&
352 buf->buf_size < JOURNAL_ENTRY_SIZE_MAX)
353 j->buf_size_want = max(j->buf_size_want, buf->buf_size << 1);
355 if (journal_entry_is_open(j) &&
356 !__journal_entry_close(j)) {
358 * We failed to get a reservation on the current open journal
359 * entry because it's full, and we can't close it because
360 * there's still a previous one in flight:
362 trace_journal_entry_full(c);
365 ret = journal_entry_open(j);
368 if ((ret == -EAGAIN || ret == -ENOSPC) &&
369 !j->res_get_blocked_start)
370 j->res_get_blocked_start = local_clock() ?: 1;
372 can_discard = j->can_discard;
373 spin_unlock(&j->lock);
378 if (ret == -ENOSPC) {
379 BUG_ON(!can_discard && (flags & JOURNAL_RES_GET_RESERVED));
382 * Journal is full - can't rely on reclaim from work item due to
385 trace_journal_full(c);
387 if (!(flags & JOURNAL_RES_GET_NONBLOCK)) {
389 bch2_journal_do_discards(j);
393 if (mutex_trylock(&j->reclaim_lock)) {
394 bch2_journal_reclaim(j);
395 mutex_unlock(&j->reclaim_lock);
406 * Essentially the entry function to the journaling code. When bcachefs is doing
407 * a btree insert, it calls this function to get the current journal write.
408 * Journal write is the structure used set up journal writes. The calling
409 * function will then add its keys to the structure, queuing them for the next
412 * To ensure forward progress, the current task must not be holding any
413 * btree node write locks.
415 int bch2_journal_res_get_slowpath(struct journal *j, struct journal_res *res,
420 closure_wait_event(&j->async_wait,
421 (ret = __journal_res_get(j, res, flags)) != -EAGAIN ||
422 (flags & JOURNAL_RES_GET_NONBLOCK));
426 /* journal_preres: */
428 static bool journal_preres_available(struct journal *j,
429 struct journal_preres *res,
432 bool ret = bch2_journal_preres_get_fast(j, res, new_u64s);
435 bch2_journal_reclaim_work(&j->reclaim_work.work);
440 int __bch2_journal_preres_get(struct journal *j,
441 struct journal_preres *res,
446 closure_wait_event(&j->preres_wait,
447 (ret = bch2_journal_error(j)) ||
448 journal_preres_available(j, res, new_u64s));
452 /* journal_entry_res: */
454 void bch2_journal_entry_res_resize(struct journal *j,
455 struct journal_entry_res *res,
458 union journal_res_state state;
459 int d = new_u64s - res->u64s;
463 j->entry_u64s_reserved += d;
467 j->cur_entry_u64s = max_t(int, 0, j->cur_entry_u64s - d);
469 state = READ_ONCE(j->reservations);
471 if (state.cur_entry_offset < JOURNAL_ENTRY_CLOSED_VAL &&
472 state.cur_entry_offset > j->cur_entry_u64s) {
473 j->cur_entry_u64s += d;
475 * Not enough room in current journal entry, have to flush it:
477 __journal_entry_close(j);
479 journal_cur_buf(j)->u64s_reserved += d;
482 spin_unlock(&j->lock);
486 /* journal flushing: */
488 u64 bch2_journal_last_unwritten_seq(struct journal *j)
493 seq = journal_cur_seq(j);
494 if (j->reservations.prev_buf_unwritten)
496 spin_unlock(&j->lock);
502 * bch2_journal_open_seq_async - try to open a new journal entry if @seq isn't
503 * open yet, or wait if we cannot
505 * used by the btree interior update machinery, when it needs to write a new
506 * btree root - every journal entry contains the roots of all the btrees, so it
507 * doesn't need to bother with getting a journal reservation
509 int bch2_journal_open_seq_async(struct journal *j, u64 seq, struct closure *cl)
511 struct bch_fs *c = container_of(j, struct bch_fs, journal);
517 * Can't try to open more than one sequence number ahead:
519 BUG_ON(journal_cur_seq(j) < seq && !journal_entry_is_open(j));
521 if (journal_cur_seq(j) > seq ||
522 journal_entry_is_open(j)) {
523 spin_unlock(&j->lock);
527 if (journal_cur_seq(j) < seq &&
528 !__journal_entry_close(j)) {
529 /* haven't finished writing out the previous one: */
530 trace_journal_entry_full(c);
533 BUG_ON(journal_cur_seq(j) != seq);
535 ret = journal_entry_open(j);
538 if ((ret == -EAGAIN || ret == -ENOSPC) &&
539 !j->res_get_blocked_start)
540 j->res_get_blocked_start = local_clock() ?: 1;
542 if (ret == -EAGAIN || ret == -ENOSPC)
543 closure_wait(&j->async_wait, cl);
545 spin_unlock(&j->lock);
547 if (ret == -ENOSPC) {
548 trace_journal_full(c);
549 bch2_journal_reclaim_work(&j->reclaim_work.work);
556 static int journal_seq_error(struct journal *j, u64 seq)
558 union journal_res_state state = READ_ONCE(j->reservations);
560 if (seq == journal_cur_seq(j))
561 return bch2_journal_error(j);
563 if (seq + 1 == journal_cur_seq(j) &&
564 !state.prev_buf_unwritten &&
571 static inline struct journal_buf *
572 journal_seq_to_buf(struct journal *j, u64 seq)
574 /* seq should be for a journal entry that has been opened: */
575 BUG_ON(seq > journal_cur_seq(j));
576 BUG_ON(seq == journal_cur_seq(j) &&
577 j->reservations.cur_entry_offset == JOURNAL_ENTRY_CLOSED_VAL);
579 if (seq == journal_cur_seq(j))
580 return journal_cur_buf(j);
581 if (seq + 1 == journal_cur_seq(j) &&
582 j->reservations.prev_buf_unwritten)
583 return journal_prev_buf(j);
588 * bch2_journal_wait_on_seq - wait for a journal entry to be written
590 * does _not_ cause @seq to be written immediately - if there is no other
591 * activity to cause the relevant journal entry to be filled up or flushed it
592 * can wait for an arbitrary amount of time (up to @j->write_delay_ms, which is
595 void bch2_journal_wait_on_seq(struct journal *j, u64 seq,
596 struct closure *parent)
598 struct journal_buf *buf;
602 if ((buf = journal_seq_to_buf(j, seq))) {
603 if (!closure_wait(&buf->wait, parent))
606 if (seq == journal_cur_seq(j)) {
608 if (bch2_journal_error(j))
609 closure_wake_up(&buf->wait);
613 spin_unlock(&j->lock);
617 * bch2_journal_flush_seq_async - wait for a journal entry to be written
619 * like bch2_journal_wait_on_seq, except that it triggers a write immediately if
622 void bch2_journal_flush_seq_async(struct journal *j, u64 seq,
623 struct closure *parent)
625 struct journal_buf *buf;
630 (buf = journal_seq_to_buf(j, seq)))
631 if (!closure_wait(&buf->wait, parent))
634 if (seq == journal_cur_seq(j))
635 __journal_entry_close(j);
636 spin_unlock(&j->lock);
639 static int journal_seq_flushed(struct journal *j, u64 seq)
644 ret = seq <= j->seq_ondisk ? 1 : journal_seq_error(j, seq);
646 if (seq == journal_cur_seq(j))
647 __journal_entry_close(j);
648 spin_unlock(&j->lock);
653 int bch2_journal_flush_seq(struct journal *j, u64 seq)
655 u64 start_time = local_clock();
658 ret = wait_event_killable(j->wait, (ret2 = journal_seq_flushed(j, seq)));
660 bch2_time_stats_update(j->flush_seq_time, start_time);
662 return ret ?: ret2 < 0 ? ret2 : 0;
666 * bch2_journal_meta_async - force a journal entry to be written
668 void bch2_journal_meta_async(struct journal *j, struct closure *parent)
670 struct journal_res res;
672 memset(&res, 0, sizeof(res));
674 bch2_journal_res_get(j, &res, jset_u64s(0), 0);
675 bch2_journal_res_put(j, &res);
677 bch2_journal_flush_seq_async(j, res.seq, parent);
680 int bch2_journal_meta(struct journal *j)
682 struct journal_res res;
685 memset(&res, 0, sizeof(res));
687 ret = bch2_journal_res_get(j, &res, jset_u64s(0), 0);
691 bch2_journal_res_put(j, &res);
693 return bch2_journal_flush_seq(j, res.seq);
697 * bch2_journal_flush_async - if there is an open journal entry, or a journal
698 * still being written, write it and wait for the write to complete
700 void bch2_journal_flush_async(struct journal *j, struct closure *parent)
702 u64 seq, journal_seq;
705 journal_seq = journal_cur_seq(j);
707 if (journal_entry_is_open(j)) {
709 } else if (journal_seq) {
710 seq = journal_seq - 1;
712 spin_unlock(&j->lock);
715 spin_unlock(&j->lock);
717 bch2_journal_flush_seq_async(j, seq, parent);
720 int bch2_journal_flush(struct journal *j)
722 u64 seq, journal_seq;
725 journal_seq = journal_cur_seq(j);
727 if (journal_entry_is_open(j)) {
729 } else if (journal_seq) {
730 seq = journal_seq - 1;
732 spin_unlock(&j->lock);
735 spin_unlock(&j->lock);
737 return bch2_journal_flush_seq(j, seq);
740 /* block/unlock the journal: */
742 void bch2_journal_unblock(struct journal *j)
746 spin_unlock(&j->lock);
751 void bch2_journal_block(struct journal *j)
755 spin_unlock(&j->lock);
760 /* allocate journal on a device: */
762 static int __bch2_set_nr_journal_buckets(struct bch_dev *ca, unsigned nr,
763 bool new_fs, struct closure *cl)
765 struct bch_fs *c = ca->fs;
766 struct journal_device *ja = &ca->journal;
767 struct bch_sb_field_journal *journal_buckets;
768 u64 *new_bucket_seq = NULL, *new_buckets = NULL;
771 /* don't handle reducing nr of buckets yet: */
776 new_buckets = kzalloc(nr * sizeof(u64), GFP_KERNEL);
777 new_bucket_seq = kzalloc(nr * sizeof(u64), GFP_KERNEL);
778 if (!new_buckets || !new_bucket_seq)
781 journal_buckets = bch2_sb_resize_journal(&ca->disk_sb,
782 nr + sizeof(*journal_buckets) / sizeof(u64));
783 if (!journal_buckets)
787 * We may be called from the device add path, before the new device has
788 * actually been added to the running filesystem:
791 spin_lock(&c->journal.lock);
793 memcpy(new_buckets, ja->buckets, ja->nr * sizeof(u64));
794 memcpy(new_bucket_seq, ja->bucket_seq, ja->nr * sizeof(u64));
795 swap(new_buckets, ja->buckets);
796 swap(new_bucket_seq, ja->bucket_seq);
799 spin_unlock(&c->journal.lock);
801 while (ja->nr < nr) {
802 struct open_bucket *ob = NULL;
807 bucket = bch2_bucket_alloc_new_fs(ca);
813 ob = bch2_bucket_alloc(c, ca, RESERVE_ALLOC,
816 ret = cl ? -EAGAIN : -ENOSPC;
820 bucket = sector_to_bucket(ca, ob->ptr.offset);
824 percpu_down_read(&c->mark_lock);
825 spin_lock(&c->journal.lock);
828 pos = ja->nr ? (ja->cur_idx + 1) % ja->nr : 0;
829 __array_insert_item(ja->buckets, ja->nr, pos);
830 __array_insert_item(ja->bucket_seq, ja->nr, pos);
831 __array_insert_item(journal_buckets->buckets, ja->nr, pos);
834 ja->buckets[pos] = bucket;
835 ja->bucket_seq[pos] = 0;
836 journal_buckets->buckets[pos] = cpu_to_le64(bucket);
838 if (pos <= ja->discard_idx)
839 ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
840 if (pos <= ja->dirty_idx_ondisk)
841 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
842 if (pos <= ja->dirty_idx)
843 ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
844 if (pos <= ja->cur_idx)
845 ja->cur_idx = (ja->cur_idx + 1) % ja->nr;
847 bch2_mark_metadata_bucket(c, ca, bucket, BCH_DATA_JOURNAL,
849 gc_phase(GC_PHASE_SB),
853 spin_unlock(&c->journal.lock);
854 percpu_up_read(&c->mark_lock);
858 bch2_open_bucket_put(c, ob);
863 kfree(new_bucket_seq);
870 * Allocate more journal space at runtime - not currently making use if it, but
873 int bch2_set_nr_journal_buckets(struct bch_fs *c, struct bch_dev *ca,
876 struct journal_device *ja = &ca->journal;
881 closure_init_stack(&cl);
884 struct disk_reservation disk_res = { 0, 0 };
888 mutex_lock(&c->sb_lock);
892 * note: journal buckets aren't really counted as _sectors_ used yet, so
893 * we don't need the disk reservation to avoid the BUG_ON() in buckets.c
894 * when space used goes up without a reservation - but we do need the
895 * reservation to ensure we'll actually be able to allocate:
898 if (bch2_disk_reservation_get(c, &disk_res,
899 bucket_to_sector(ca, nr - ja->nr), 1, 0)) {
900 mutex_unlock(&c->sb_lock);
904 ret = __bch2_set_nr_journal_buckets(ca, nr, false, &cl);
906 bch2_disk_reservation_put(c, &disk_res);
908 if (ja->nr != current_nr)
910 mutex_unlock(&c->sb_lock);
911 } while (ret == -EAGAIN);
916 int bch2_dev_journal_alloc(struct bch_dev *ca)
920 if (dynamic_fault("bcachefs:add:journal_alloc"))
924 * clamp journal size to 1024 buckets or 512MB (in sectors), whichever
927 nr = clamp_t(unsigned, ca->mi.nbuckets >> 8,
928 BCH_JOURNAL_BUCKETS_MIN,
930 (1 << 20) / ca->mi.bucket_size));
932 return __bch2_set_nr_journal_buckets(ca, nr, true, NULL);
935 /* startup/shutdown: */
937 static bool bch2_journal_writing_to_device(struct journal *j, unsigned dev_idx)
939 union journal_res_state state;
940 struct journal_buf *w;
944 state = READ_ONCE(j->reservations);
945 w = j->buf + !state.idx;
947 ret = state.prev_buf_unwritten &&
948 bch2_extent_has_device(bkey_i_to_s_c_extent(&w->key), dev_idx);
949 spin_unlock(&j->lock);
954 void bch2_dev_journal_stop(struct journal *j, struct bch_dev *ca)
956 wait_event(j->wait, !bch2_journal_writing_to_device(j, ca->dev_idx));
959 void bch2_fs_journal_stop(struct journal *j)
961 struct bch_fs *c = container_of(j, struct bch_fs, journal);
963 bch2_journal_flush_all_pins(j);
965 wait_event(j->wait, journal_entry_close(j));
967 /* do we need to write another journal entry? */
968 if (test_bit(JOURNAL_NOT_EMPTY, &j->flags) ||
969 c->btree_roots_dirty)
970 bch2_journal_meta(j);
974 BUG_ON(!bch2_journal_error(j) &&
975 test_bit(JOURNAL_NOT_EMPTY, &j->flags));
977 cancel_delayed_work_sync(&j->write_work);
978 cancel_delayed_work_sync(&j->reclaim_work);
981 int bch2_fs_journal_start(struct journal *j, u64 cur_seq,
982 struct list_head *journal_entries)
984 struct bch_fs *c = container_of(j, struct bch_fs, journal);
985 struct journal_entry_pin_list *p;
986 struct journal_replay *i;
987 u64 last_seq = cur_seq, nr, seq;
989 if (!list_empty(journal_entries))
990 last_seq = le64_to_cpu(list_first_entry(journal_entries,
991 struct journal_replay,
994 nr = cur_seq - last_seq;
996 if (nr + 1 > j->pin.size) {
998 init_fifo(&j->pin, roundup_pow_of_two(nr + 1), GFP_KERNEL);
1000 bch_err(c, "error reallocating journal fifo (%llu open entries)", nr);
1005 j->replay_journal_seq = last_seq;
1006 j->replay_journal_seq_end = cur_seq;
1007 j->last_seq_ondisk = last_seq;
1008 j->pin.front = last_seq;
1009 j->pin.back = cur_seq;
1010 atomic64_set(&j->seq, cur_seq - 1);
1012 fifo_for_each_entry_ptr(p, &j->pin, seq) {
1013 INIT_LIST_HEAD(&p->list);
1014 INIT_LIST_HEAD(&p->flushed);
1015 atomic_set(&p->count, 1);
1019 list_for_each_entry(i, journal_entries, list) {
1020 seq = le64_to_cpu(i->j.seq);
1022 BUG_ON(seq < last_seq || seq >= cur_seq);
1024 journal_seq_pin(j, seq)->devs = i->devs;
1027 spin_lock(&j->lock);
1029 set_bit(JOURNAL_STARTED, &j->flags);
1031 journal_pin_new_entry(j, 1);
1032 bch2_journal_buf_init(j);
1034 c->last_bucket_seq_cleanup = journal_cur_seq(j);
1036 bch2_journal_space_available(j);
1037 spin_unlock(&j->lock);
1044 void bch2_dev_journal_exit(struct bch_dev *ca)
1046 kfree(ca->journal.bio);
1047 kfree(ca->journal.buckets);
1048 kfree(ca->journal.bucket_seq);
1050 ca->journal.bio = NULL;
1051 ca->journal.buckets = NULL;
1052 ca->journal.bucket_seq = NULL;
1055 int bch2_dev_journal_init(struct bch_dev *ca, struct bch_sb *sb)
1057 struct journal_device *ja = &ca->journal;
1058 struct bch_sb_field_journal *journal_buckets =
1059 bch2_sb_get_journal(sb);
1062 ja->nr = bch2_nr_journal_buckets(journal_buckets);
1064 ja->bucket_seq = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1065 if (!ja->bucket_seq)
1068 ca->journal.bio = bio_kmalloc(GFP_KERNEL,
1069 DIV_ROUND_UP(JOURNAL_ENTRY_SIZE_MAX, PAGE_SIZE));
1070 if (!ca->journal.bio)
1073 ja->buckets = kcalloc(ja->nr, sizeof(u64), GFP_KERNEL);
1077 for (i = 0; i < ja->nr; i++)
1078 ja->buckets[i] = le64_to_cpu(journal_buckets->buckets[i]);
1083 void bch2_fs_journal_exit(struct journal *j)
1085 kvpfree(j->buf[1].data, j->buf[1].buf_size);
1086 kvpfree(j->buf[0].data, j->buf[0].buf_size);
1090 int bch2_fs_journal_init(struct journal *j)
1092 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1093 static struct lock_class_key res_key;
1096 pr_verbose_init(c->opts, "");
1098 spin_lock_init(&j->lock);
1099 spin_lock_init(&j->err_lock);
1100 init_waitqueue_head(&j->wait);
1101 INIT_DELAYED_WORK(&j->write_work, journal_write_work);
1102 INIT_DELAYED_WORK(&j->reclaim_work, bch2_journal_reclaim_work);
1103 init_waitqueue_head(&j->pin_flush_wait);
1104 mutex_init(&j->reclaim_lock);
1105 mutex_init(&j->discard_lock);
1107 lockdep_init_map(&j->res_map, "journal res", &res_key, 0);
1109 j->buf[0].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1110 j->buf[1].buf_size = JOURNAL_ENTRY_SIZE_MIN;
1111 j->write_delay_ms = 1000;
1112 j->reclaim_delay_ms = 100;
1115 j->entry_u64s_reserved +=
1116 BTREE_ID_NR * (JSET_KEYS_U64s + BKEY_EXTENT_U64s_MAX);
1118 atomic64_set(&j->reservations.counter,
1119 ((union journal_res_state)
1120 { .cur_entry_offset = JOURNAL_ENTRY_CLOSED_VAL }).v);
1122 if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
1123 !(j->buf[0].data = kvpmalloc(j->buf[0].buf_size, GFP_KERNEL)) ||
1124 !(j->buf[1].data = kvpmalloc(j->buf[1].buf_size, GFP_KERNEL))) {
1129 j->pin.front = j->pin.back = 1;
1131 pr_verbose_init(c->opts, "ret %i", ret);
1137 ssize_t bch2_journal_print_debug(struct journal *j, char *buf)
1139 struct printbuf out = _PBUF(buf, PAGE_SIZE);
1140 struct bch_fs *c = container_of(j, struct bch_fs, journal);
1141 union journal_res_state s;
1146 spin_lock(&j->lock);
1147 s = READ_ONCE(j->reservations);
1150 "active journal entries:\t%llu\n"
1152 "last_seq:\t\t%llu\n"
1153 "last_seq_ondisk:\t%llu\n"
1154 "prereserved:\t\t%u/%u\n"
1155 "current entry sectors:\t%u\n"
1156 "current entry:\t\t",
1159 journal_last_seq(j),
1161 j->prereserved.reserved,
1162 j->prereserved.remaining,
1163 j->cur_entry_sectors);
1165 switch (s.cur_entry_offset) {
1166 case JOURNAL_ENTRY_ERROR_VAL:
1167 pr_buf(&out, "error\n");
1169 case JOURNAL_ENTRY_CLOSED_VAL:
1170 pr_buf(&out, "closed\n");
1173 pr_buf(&out, "%u/%u\n",
1180 "current entry refs:\t%u\n"
1181 "prev entry unwritten:\t",
1182 journal_state_count(s, s.idx));
1184 if (s.prev_buf_unwritten)
1185 pr_buf(&out, "yes, ref %u sectors %u\n",
1186 journal_state_count(s, !s.idx),
1187 journal_prev_buf(j)->sectors);
1189 pr_buf(&out, "no\n");
1192 "need write:\t\t%i\n"
1193 "replay done:\t\t%i\n",
1194 test_bit(JOURNAL_NEED_WRITE, &j->flags),
1195 test_bit(JOURNAL_REPLAY_DONE, &j->flags));
1197 for_each_member_device_rcu(ca, c, iter,
1198 &c->rw_devs[BCH_DATA_JOURNAL]) {
1199 struct journal_device *ja = &ca->journal;
1207 "\tavailable\t%u:%u\n"
1208 "\tdiscard_idx\t\t%u\n"
1209 "\tdirty_idx_ondisk\t%u (seq %llu)\n"
1210 "\tdirty_idx\t\t%u (seq %llu)\n"
1211 "\tcur_idx\t\t%u (seq %llu)\n",
1213 bch2_journal_dev_buckets_available(j, ja, journal_space_discarded),
1216 ja->dirty_idx_ondisk, ja->bucket_seq[ja->dirty_idx_ondisk],
1217 ja->dirty_idx, ja->bucket_seq[ja->dirty_idx],
1218 ja->cur_idx, ja->bucket_seq[ja->cur_idx]);
1221 spin_unlock(&j->lock);
1224 return out.pos - buf;
1227 ssize_t bch2_journal_print_pins(struct journal *j, char *buf)
1229 struct printbuf out = _PBUF(buf, PAGE_SIZE);
1230 struct journal_entry_pin_list *pin_list;
1231 struct journal_entry_pin *pin;
1234 spin_lock(&j->lock);
1235 fifo_for_each_entry_ptr(pin_list, &j->pin, i) {
1236 pr_buf(&out, "%llu: count %u\n",
1237 i, atomic_read(&pin_list->count));
1239 list_for_each_entry(pin, &pin_list->list, list)
1240 pr_buf(&out, "\t%p %pf\n",
1243 if (!list_empty(&pin_list->flushed))
1244 pr_buf(&out, "flushed:\n");
1246 list_for_each_entry(pin, &pin_list->flushed, list)
1247 pr_buf(&out, "\t%p %pf\n",
1250 spin_unlock(&j->lock);
1252 return out.pos - buf;